Aare provides a Python library that adds bidirectional guardrails to LLM applications, validating inputs for injection attacks and protected health information and formally verifying outputs with the Z3 theorem prover to ensure no PHI, credit‑card data, or policy violations are present. The framework integrates with LangChain and other LLM toolkits, offering configurable block, warn, or redact actions and can run in-browser or on‑premises for zero‑trust compliance in regulated industries.
Funding
Funding not disclosed
Founders
Product
Problem
Organizations using large language models often rely on fragile regex filters or single‑direction guardrails, leaving them vulnerable to prompt injection attacks, leakage of protected health information (PHI), and violations of regulations such as HIPAA, PCI DSS, GDPR, and corporate policies.
Solution
Aare offers a full‑pipeline guardrail framework that validates user inputs before they reach an LLM and formally verifies LLM outputs before they are returned to users. The input guardrail detects injection attempts and PHI in prompts, while the output guardrail uses Z3 theorem‑proving to generate a formal proof that no protected data or policy violations are present. The system can be configured to block, warn, or redact offending content, providing deterministic safety without relying on prompt engineering. Aare’s library integrates with LangChain and other LLM toolkits via a simple Python API, enabling developers to embed provable compliance into any generative AI application. By running entirely in the browser or on‑premises, it eliminates the need to send sensitive data to external services.
Target Audience
Primary customers are enterprises and regulated industries—such as healthcare, finance, and legal services—that build or operate LLM‑powered applications and need provable compliance with data protection regulations.
Features
- Input guardrail that scans prompts for injection patterns and all 18 HIPAA Safe Harbor PHI categories
- Output guardrail that performs formal verification with the Z3 theorem prover to ensure no PHI, credit‑card data, or policy‑restricted content is generated
- Post‑generation immunity to jailbreaks, preventing prompt injection after the LLM has responded
- Configurable actions (block, warn, redact) for both input and output violations
- Seamless Python integration (e.g., `input_guard | prompt | llm | output_guard`) compatible with LangChain and other LLM frameworks
- Browser‑based demo and on‑premises deployment options for zero‑trust environments